On 7/25/21 6:27 PM, Vlastimil Babka wrote:
>>> Hope fixing that helps then?
>>
>> Yeah, though RT should perhaps be pinned across release/re-acquire?
>
> Probably not necessary, this_cpu_cmpxchg() will effectively recognize
> being moved to a different CPU.
> Might also move __unfreeze_partials() out of the whole loop to avoid the
> relock. Yeah that should be better.
>
>> Actually, local locks should rediscover the recursion handling skills
>> they long had so such RT specific hole poking isn't necessary. There
>> previously would have been no ifdef+typo there for eyeballs to miss and
>> miss and miss.
>
> Hm, now I'm realizing that local_lock() on !RT is just
> preempt_disable(), i.e. equivalent to get_cpu_ptr(), so some of the
> ifdeffery could go away?
How much will this explode? Thanks.
----8<----
>From 99808b198bdf867951131bb9d1ca1bd1cd12b8c4 Mon Sep 17 00:00:00 2001
From: Vlastimil Babka <vbabka@xxxxxxx>
Date: Fri, 23 Jul 2021 23:17:18 +0200
Subject: [PATCH] PREEMPT_RT+SLUB_CPU_PARTIAL fix attempt
---
mm/slub.c | 34 +++++++++++++++++++++++++++++++---
1 file changed, 31 insertions(+), 3 deletions(-)
diff --git a/mm/slub.c b/mm/slub.c
index 581004a5aca9..d12a50b5ee6f 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2437,13 +2437,19 @@ static void unfreeze_partials(struct kmem_cache *s)
{
struct page *partial_page;
+#ifndef CONFIG_PREEMPT_RT
do {
partial_page = this_cpu_read(s->cpu_slab->partial);
} while (partial_page &&
this_cpu_cmpxchg(s->cpu_slab->partial, partial_page, NULL)
!= partial_page);
-
+#else
+ local_lock(&s->cpu_slab->lock);
+ partial_page = this_cpu_read(s->cpu_slab->partial);
+ this_cpu_write(s->cpu_slab->partial, NULL);
+ local_unlock(&s->cpu_slab->lock);
+#endif
if (partial_page)
__unfreeze_partials(s, partial_page);
}
@@ -2479,10 +2485,15 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
{
#ifdef CONFIG_SLUB_CPU_PARTIAL
struct page *oldpage;
+ struct page *page_to_unfreeze = NULL;
int pages;
int pobjects;
- slub_get_cpu_ptr(s->cpu_slab);
+ /*
+ * On !RT we just want to disable preemption, on RT we need the lock
+ * for real. This happens to match local_lock() semantics.
+ */
+ local_lock(&s->cpu_slab->lock);
do {
pages = 0;
pobjects = 0;
@@ -2496,7 +2507,16 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
* partial array is full. Move the existing
* set to the per node partial list.
*/
+#ifndef CONFIG_PREEMPT_RT
unfreeze_partials(s);
+#else
+ /*
+ * Postpone unfreezing until we drop the local
+ * lock to avoid relocking.
+ */
+ page_to_unfreeze = oldpage;
+#endif
+
oldpage = NULL;
pobjects = 0;
pages = 0;
@@ -2511,9 +2531,17 @@ static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
page->pobjects = pobjects;
page->next = oldpage;
+#ifndef CONFIG_PREEMPT_RT
} while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
!= oldpage);
- slub_put_cpu_ptr(s->cpu_slab);
+#else
+ this_cpu_write(s->cpu_slab->partial, page);
+ } while (false);
+#endif
+
+ local_unlock(&s->cpu_slab->lock);
+ if (page_to_unfreeze)
+ __unfreeze_partials(s, page_to_unfreeze);
#endif /* CONFIG_SLUB_CPU_PARTIAL */
}
--
2.32.0
